Fooling the Big Picture in Classification Tasks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• One-stage classifiers are susceptible to adversarial examples that reduce performance while remaining imperceptible.
• Hierarchical classifiers utilize fine and coarse level categories, presenting a different attack surface.

Purpose of the Study:

• Investigate the susceptibility of hierarchical classifiers to adversarial attacks.
• Develop methods to induce misclassification at the coarse level of hierarchical classifiers.
• Analyze the trade-offs between attack effectiveness and imperceptibility.

Main Methods:

• Formulated a minimax optimization program to encode adversarial attack constraints.
• Developed solutions using convex relaxations of the optimization program.
• Employed the alternating direction method of multipliers (ADMM) for algorithm development.

Main Results:

• Successfully fooled coarse classification in hierarchical models, demonstrated by relative accuracy loss.
• Attacks targeting the 'big picture' (coarse classification) required higher perturbation levels compared to one-stage classifiers.
• Achieved competitive performance against state-of-the-art solvers.

Conclusions:

• Hierarchical classifiers exhibit susceptibility to adversarial attacks targeting coarse-level predictions.
• Attacking coarse categories in hierarchical classifiers necessitates greater perturbation, leading to reduced imperceptibility.
• The impact of label groupings on attack performance was also examined.